Parameter-Efficient Federated Adaptive Fine-Tuning for Fault Diagnosis Under Heterogeneous Measurement Data

Effective and efficient sensor data analysis is crucial for achieving scientific and reliable machinery fault diagnosis. Recently, federated learning (FL) has been widely studied to collaboratively train fault diagnosis models on the distributed sensor data while protecting data privacy. Nevertheless, the performance of existing FL methods is often hampered by intraclient and interclient data heterogeneity. Moreover, they typically require all model parameters to be transmitted between the server and clients, resulting in substantial communication overhead. To address these challenges, a novel parameter-efficient federated adaptive fine-tuning framework is developed. First, the label-smoothing strategy is introduced to facilitate collaborative model pre-training across multiple clients. Then, a lightweight feature scaling and shifting pool is designed to adjust the features extracted by the pre-trained model through federated fine-tuning. Finally, to make the global model adapt to the heterogeneous measurement data of different clients, an instance-adaptive fault diagnosis mechanism is developed to set customized tuning parameters for the model based on the features of each test sample. Extensive experiments conducted on two self-built test platforms show that the proposed method improves diagnosis accuracy by 3.47% and reduces communication cost by 92.40% compared with state-of-the-art FL methods. These results demonstrate the superiority and practicality of our method in heterogeneous industrial environments.